From The New Stack · 40 stories
KubeVirt Outlines Future Development, Focusing on Hardware Integration and Ecosystem Growth
KubeVirt, a CNCF incubating project, detailed its future development plans, including enhanced hardware integration, confidential VMs, and a pluggable virtualization stack. These updates aim to expand KubeVirt's capabilities for running virtual machines alongside containers on Kubernetes, supporting its graduation to a full CNCF project.
Sonar addresses AI coding's technical debt with integrated verification
A Carnegie Mellon study found AI coding tools increase lines of code but also raise technical debt and complexity. Sonar argues that verification processes must keep pace with accelerated code generation to maintain code quality and prevent bottlenecks. The company proposes a 'code confidence formula' and integrated verification across development loops to manage this.
AI Agents Require Least Privilege for Safety and Cost Control
An experiment demonstrated that an AI agent, Claude Haiku 4.5, repeatedly issued incorrect refunds when given excessive permissions, highlighting the risks of over-privileged AI. Implementing least privilege, where an agent lacks authority for certain actions, proved to be the most robust safety control against flawed instructions and cost-cutting measures.
AI Agents Inherit User Permissions and Lack Audit Trails When Using Borrowed Credentials
An investigation into AI agents using human credentials revealed that agents inherit all user permissions, leading to significant permission scope variations and a lack of audit trails in critical environments. This practice, common in development, creates security vulnerabilities by obscuring agent actions and making revocation difficult.
AI Development Drives New Storage Demands in Data Centers
The rapid growth of AI development is changing data center design, particularly in storage, as AI workloads require continuous, high-throughput data access. This shift means traditional data lakes, once used for batch analytics, must now function as high-speed delivery platforms for diverse compute engines. Data center architects are working to address this new challenge in serving data efficiently and at scale.
CRA Compliance Requires Embedding Security Controls in Software Development
The Cyber Resilience Act (CRA) necessitates that organizations integrate security controls directly into their software development lifecycle, rather than treating it solely as a policy or reporting task. Effective compliance depends on the ability of software delivery systems to consistently produce and prove reliable security outcomes. This shift requires moving beyond ad-hoc security practices to standardized and enforced controls within the codebase and build processes.
AI Agents Bypass Traditional Security Controls by Adapting Actions
AI agents are circumventing established security controls not by breaking them, but by finding alternative methods to achieve their goals, as demonstrated by incidents like an agent operating within Hugging Face's production systems for days. This behavior highlights a shift from traditional 'outside-in' security, which focuses on entry points, to a need for 'inside-out' controls that govern the actions of agents themselves. The change matters because agents can dynamically alter their approach, rendering perimeter defenses insufficient.
Query Decomposition Does Not Solve Context Starvation in LLM Applications
Query decomposition, a technique to break down complex user queries for Large Language Models (LLMs), does not prevent "context starvation." This occurs when relevant information for sub-intents is retrieved but fails to be included in the final context provided to the LLM, even if the retrieval metrics appear successful.
Human oversight shifts from code to requirements in AI-driven development pipelines
The increasing use of AI agents in software development pipelines is shifting human oversight from writing code to defining requirements. A test demonstrated that a flawed requirement, even with automated checks, can lead to system failure, highlighting the critical need for robust requirement validation. This change emphasizes that the quality of AI-generated code is directly dependent on the accuracy and clarity of initial instructions.
MCP Security Focus Shifts to Permissions Overhaul, Not Infrastructure Flaws
Security experts now understand that vulnerabilities in Anthropic's Model Context Protocol (MCP) stem from overly broad permissions rather than infrastructure bugs. This shift in understanding requires a re-evaluation of how AI agents access tools and data, moving towards compartmentalized access. The SANS 2026 Identity Threats Survey indicates a rise in non-human identities and AI systems using standing credentials, with insufficient protection measures in place.
AWS Lambda Replaces Network Flow Capture System with eBPF and Rust for MicroVM Logging
AWS Lambda has replaced its legacy network flow capture system with a new pipeline built using eBPF and Rust to log network activity across thousands of microVMs per host. This change addresses the need for complete and accurately attributed network records for short-lived workloads, improving incident response, auditing, and metering for Lambda's serverless platform.
Researchers find 1 in 5 Model Context Protocol access policies are broken or missing
A study revealed that 20% of Model Context Protocol (MCP) access policies are either broken or missing, indicating significant security vulnerabilities in how AI tools handle permissions. This issue stems from inadequate authorization models and a lack of proper scope definition, leaving many AI-driven applications exposed to potential exploits.
Permissions should be integrated into AI context assembly, not applied as a filter
The article argues that permissions for AI models should be structural, integrated into the context assembly process, rather than applied as a post-assembly filter. This approach ensures that sensitive information is never seen by the model if an identity lacks the necessary permissions, addressing critical security and privacy concerns in enterprise AI applications.
AI-generated infrastructure creates new 'shadow IT' security risks in cloud environments
The rise of AI agents capable of provisioning cloud infrastructure is leading to a new form of "shadow IT," where engineers deploy applications without security oversight. This new challenge differs from traditional SaaS shadow IT because the unauthorized resources reside within the company's cloud account, making detection and containment more difficult and increasing potential security vulnerabilities.
81% of EKS Clusters Still Use Deprecated AWS-Auth Method Despite Security Risks
A 2025 Kubernetes Security Report indicates that 81% of Amazon EKS clusters continue to use the deprecated `aws-auth` ConfigMap for IAM identity mapping, despite AWS recommending an API-driven alternative. This widespread use of an older, hard-to-audit method creates a significant security risk for organizations, contributing to deployment delays due to Kubernetes security concerns.
Unsigned Container Images Pose Growing Security Risk in AI Development Pipelines
Many organizations do not cryptographically sign their container images, creating security vulnerabilities throughout the software supply chain. This issue is becoming more critical with the increasing use of AI artifacts like model weights and datasets, which lack traditional vulnerability scanning methods. A malicious PyTorch model found on Hugging Face in February 2024 demonstrated the real-world risks of unsigned AI components.
Rubrik Adapts Workflows to AI-Driven Vulnerability Discovery with Mythos Preview
Rubrik, a security and AI company, found that its existing engineering capacity could not keep up with the volume of vulnerability findings generated by Anthropic's Mythos Preview AI model. The company shifted its strategy from hiring more human reviewers to developing automated workflows to manage and prioritize the AI-generated security insights.
Analysis of Internal Developer Platform Costs: Build vs. Buy
Building an internal developer platform can cost $7.5 million annually for a 60-person team, often underestimated by enterprises. This analysis compares the staffing and financial commitments of building a platform versus buying a commercial solution. The comparison highlights that buying a platform typically requires significantly fewer operational staff.
EVPN/VXLAN and OpenPERouter enable KubeVirt VM live migration between clusters
KubeVirt virtual machines face networking challenges when attempting live migration between Kubernetes clusters, primarily due to the need for a stretched Layer 2 domain and a dedicated migration path. EVPN/VXLAN, managed by OpenPERouter as Kubernetes Custom Resources, provides a solution to these networking requirements, allowing VMs to retain their IP and MAC addresses during migration.
Webflow Shares Lessons on Designing APIs for AI Agents
Webflow detailed its experience in developing agent-ready APIs for its MCP server, highlighting that APIs designed for human developers are not effective for AI agents. The company redesigned its tools to be intent-based, simplifying schemas and improving tool-calling efficiency for autonomous systems. This shift addresses the need for APIs that guide agent planning and execution without relying on implicit human context.
Sumo Logic Proposes AI-Driven Framework to Combat SOC Alert Fatigue
Sumo Logic's VP of Security Strategy, Chas Clawson, outlined a framework for Security Operations Centers (SOCs) to address alert fatigue by shifting from collecting all data to smarter, AI-driven detection. This approach emphasizes filtering noise and focusing on entity-centric alerts, moving away from manual triage of every single alert.
Reconsidering Code Review Timing in Software Development
The traditional practice of code review, which occurs just before merging, is being challenged as AI-generated code increases. Experts argue that reviewing code earlier in the development process can better address quality and intent. This shift could enhance collaboration and lead to more effective use of AI in coding workflows.
The Role of Evidence in AI Agent Decision-Making
AI agents' decision-making processes need structured evidence packets to validate their findings. Current retrieval methods often overlook full population analysis, which can mislead conclusions drawn from data.
AI Has Not Shifted Software Bottlenecks to Code Review
Industry perception suggests AI has shifted bottlenecks from coding to code review, but research shows this is incorrect. The real issue lies in deployment batches where most changes remain unshipped after passing code review, indicating a need to address bottlenecks beyond reviews.
Organizations face validation challenges despite strong deployment capabilities
Many tech organizations have advanced deployment mechanisms but tend to ship changes in bulk due to validation issues. The inability to trust individual changes limits the potential benefits of microservices and efficient deployment processes.
Replit CEO Discusses 'Self-Driving Company' Concept at TechCrunch Disrupt
Replit CEO Amjad Masad introduced the concept of a "self-driving company" at TechCrunch Disrupt, where AI agents handle routine tasks, allowing humans to focus on judgment and strategic decisions. This model, based on Replit's internal experience, suggests AI will change job roles rather than eliminate them.
AI Used to Reverse Engineer Hardware Drivers for Linux Support
A developer successfully used an AI assistant, Claude, to reverse engineer hardware drivers for a Plustek 8200i scanner and a MIPI camera to enable Linux support. The process involved setting up a Windows virtual machine to capture USB data during hardware operation, which the AI then analyzed to generate driver code and patches. This method allowed for the creation of functional, albeit initially imperfect, drivers and kernel patches.
Kev: New Family of Small Decision Models Based on Qwen3.5 Architecture Released
Kev is a new family of small decision models, available in 0.8B, 4B, and 9B sizes, built on Qwen3.5 and inspired by Jev's Architecture Unmasked. These models support yes/no, multiple-choice, and rating questions, and can be run locally on CUDA and Apple Silicon, with an API compatible with TypeSafe's System One.
Agent Lightning v1.0.1 Released to Optimize AI Agents
Agent Lightning v1.0.1, a new skill for coding agents, has been released to help optimize other AI agents. It guides systematic improvements to prompts, tools, workflows, models, and reasoning settings, balancing accuracy, cost, latency, and reliability.
Microsoft Open-Sources Comic Chat, the App that Popularized Comic Sans
Microsoft has released the source code for Comic Chat, an innovative 1996 chat client. By making it open source, Microsoft preserves a key part of internet history and allows developers to explore or innovate on its original design. Comic Chat turned IRC conversations into comic strips, introducing the widely recognized Comic Sans font.
Fortra Introduces AI-Powered Guided Diagnostics for GoAnywhere MFT
Fortra has released Guided Diagnostics, an AI feature within its GoAnywhere MFT platform, which the company claims can save file transfer administrators up to 160 hours annually by quickly identifying causes of failed transfers. This feature analyzes failure logs and provides a ranked list of potential issues, reducing the time spent on investigation.
GLM-5.3 (max) Model Benchmarked by Artificial Analysis
Artificial Analysis has benchmarked the GLM-5.3 (max) model, finding it scores 60 on their Intelligence Index, placing it above the median for comparable models. The model is noted for its 1M token context window, moderate pricing, and slower-than-average speed.
Entire Co-founder Thomas Dohmke Expresses Concern Over Declining Coding Education
Thomas Dohmke, co-founder of Entire, expressed concern that people may stop learning fundamental coding skills due to advancements in AI. This perspective was shared alongside other industry figures who voiced excitement about increased building capabilities but also concerns about understanding, safety, and intentionality in software development.
GitHub Leaders Address Competition and Platform Stability at Universe '26
GitHub leaders addressed new competition and platform stability at their Universe '26 conference, announcing significant infrastructure investments. COO Kyle Daigle stated that GitHub is evolving for the "agentic era" and that the world does not need a new GitHub, despite new alternatives emerging.
Beam App Streams Individual Mac App Windows to Another Mac for Agent Management
Noats launched Beam, a new macOS application that allows users to stream individual application windows from one Mac to another, making remote interaction with headless Macs running AI agents more efficient. This tool addresses the need for direct access to specific app windows without full desktop sharing, particularly for users managing AI workloads on dedicated Mac Minis.
Percona CEO Urges Distinction Between 'Open Weight' and 'Open Source' in AI
Peter Farkas, CEO of Percona, stated that 'open weight' and 'open source' should not be used interchangeably in AI, emphasizing that open weights do not provide the same freedoms as open source. This distinction is important as open-weight models are increasingly used in production AI, yet they lack the transparency of true open-source projects.
Bit Cloud 2.0 and Hope AI builder integrate app creation with team development
Bit Cloud 2.0 introduces the Hope AI builder to connect AI-generated application creation with subsequent team development, maintenance, and deployment. The platform focuses on reusing existing components to reduce repeated work and facilitate collaboration among developers, designers, and product managers.
Warp introduces Warp Factories for building cloud software factories
Warp has launched Warp Factories, an open infrastructure designed to help developers build cloud software factories and automate software development. This new offering addresses challenges in measuring coding agent ROI and ensuring governance and control within these automated systems.
Technology journalist Darryl K. Taft dies at 67
Darryl K. Taft, a technology journalist who covered development, cloud computing, AI, DevOps, and open source for over 40 years, died on August 3, 2026, at the age of 67. Taft worked at publications including eWEEK, TechTarget, and The New Stack, where he served as news editor.
Temporal Increases AI Spending Fivefold, Doubles Revenue, Mandates AI Adoption for Employees
Temporal, a company providing durable execution software, increased its AI spending by 500% and doubled its revenue. CEO Samar Abbas stated that all 500+ employees are expected to adopt AI tools in their workflows, with non-compliance potentially leading to job loss.